Resolution generalization of deep learning-based dipole inversion networks for QSM.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Deep learning models for quantitative susceptibility mapping (QSM) struggle with varying data resolutions.
• Existing methods to improve resolution generalizability often require modifying network architecture or parameters, limiting their use with pre-trained networks.

Purpose of the Study:

• To develop a novel pipeline enabling pre-trained dipole inversion networks to reconstruct QSM from local field maps of diverse resolutions.
• To enhance the resolution generalizability of existing deep learning models for QSM without altering their architecture or parameters.

Main Methods:

• A four-step pipeline was developed: re-sampling local field maps to network-trained resolution, inferring QSM maps, combining inferred maps, and compensating for systematic errors using "dipole compensation".
• The pipeline's performance was evaluated against interpolation and naïve-input methods using the same pre-trained QSM network.

Main Results:

• The proposed pipeline demonstrated superior performance compared to alternative methods in qualitative and quantitative evaluations.
• Testing on a 1 mm³ resolution map with a network trained at 1.5 mm³ resolution showed the proposed pipeline outperformed alternatives (e.g., lower NRMSE, higher SSIM and PSNR).

Conclusions:

• The developed pipeline effectively enhances the generalizability of pre-trained dipole inversion networks to different input data resolutions.
• This approach offers a promising solution for improving the clinical applicability of deep learning-based QSM reconstruction.